Skip to contentSkip to conversation

do not try to sync identity map for non-mapped pages

PATCH In review Dave Hansen 5 participants 8 messages Mar 7, 2013 – Apr 8, 2013

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 do not try to sync identity map for non-mapped pages

The original bug reporter says this fixes it for him, so I'm broadening the cc list a bit. I assume this should just get sucked in to the x86 tree.

The double-signed-off-by from my is because my IBM email is going away very shortly.

--

kernel_map_sync_memtype() is called from a variety of contexts. The pat.c code that calls it seems to ensure that it is not called for non-ram areas by checking via pat_pagerange_is_ram(). It is important that it only be called on the actual identity map because there *IS* no map to sync for highmem pages, or for memory holes.

The ioremap.c uses are not as careful as those from pat.c, and call kernel_map_sync_memtype() on PCI space which is in the middle of the kernel identity map _range_, but is not actually mapped.

This patch adds a check to kernel_map_sync_memtype() which probably duplicates some of the checks already in pat.c. But, it is necessary for the ioremap.c uses and shouldn't hurt other callers.

I have reproduced this bug and this patch fixes it for me and the original bug reporter:

	https://lkml.org/lkml/2013/2/5/396

---

 linux-2.6.git-dave/arch/x86/mm/pat.c |    7 +++++++
 1 file changed, 7 insertions(+)
arch/x86/mm/pat.c renamed from arch/x86/mm/pat.c~dont-flush-map-for-non-ram-pages +7 -0
diff -puN arch/x86/mm/pat.c~dont-flush-map-for-non-ram-pages arch/x86/mm/pat.c
@@ -563,6 +563,13 @@ int kernel_map_sync_memtype(u64 base, un
 	if (base > __pa(high_memory-1))
 		return 0;
 
+	/*
+	 * some areas in the middle of the kernel identity range
+	 * are not mapped, like the PCI space.
+	 */
+	if (!page_is_ram(base >> PAGE_SHIFT))
+		return 0;
+
 	id_sz = (__pa(high_memory-1) <= base + size) ?
 				__pa(high_memory) - base :
 				size;

_

#1 lore raw

[tip:x86/urgent] x86: Do not try to sync identity map for non-mapped pages

x86: Do not try to sync identity map for non-mapped pages
Commit-ID:  60f583d56aa515b896a9d94f860f52640c1e8a75
Gitweb:     http://git.kernel.org/tip/60f583d56aa515b896a9d94f860f52640c1e8a75
Author:     Dave Hansen <[email protected]>
AuthorDate: Thu, 7 Mar 2013 08:31:51 -0800
Committer:  H. Peter Anvin <[email protected]>
CommitDate: Thu, 7 Mar 2013 13:23:28 -0800

x86: Do not try to sync identity map for non-mapped pages

kernel_map_sync_memtype() is called from a variety of contexts. The pat.c code that calls it seems to ensure that it is not called for non-ram areas by checking via pat_pagerange_is_ram(). It is important that it only be called on the actual identity map because there *IS* no map to sync for highmem pages, or for memory holes.

The ioremap.c uses are not as careful as those from pat.c, and call kernel_map_sync_memtype() on PCI space which is in the middle of the kernel identity map _range_, but is not actually mapped.

This patch adds a check to kernel_map_sync_memtype() which probably duplicates some of the checks already in pat.c. But, it is necessary for the ioremap.c uses and shouldn't hurt other callers.

I have reproduced this bug and this patch fixes it for me and the original bug reporter:

	https://lkml.org/lkml/2013/2/5/396
---
 arch/x86/mm/pat.c | 7 +++++++
 1 file changed, 7 insertions(+)
arch/x86/mm/pat.c +7 -0
index 2610bd9..6574388 100644
@@ -563,6 +563,13 @@ int kernel_map_sync_memtype(u64 base, unsigned long size, unsigned long flags)
 	if (base > __pa(high_memory-1))
 		return 0;
 
+	/*
+	 * some areas in the middle of the kernel identity range
+	 * are not mapped, like the PCI space.
+	 */
+	if (!page_is_ram(base >> PAGE_SHIFT))
+		return 0;
+
 	id_sz = (__pa(high_memory-1) <= base + size) ?
 				__pa(high_memory) - base :
 				size;
#2 lore raw
In reply to Dave Hansen
Quoted 20 lines · Dave Hansen

The original bug reporter says this fixes it for him, so I'm broadening the cc list a bit. I assume this should just get sucked in to the x86 tree.

The double-signed-off-by from my is because my IBM email is going away very shortly.

--

kernel_map_sync_memtype() is called from a variety of contexts. The pat.c code that calls it seems to ensure that it is not called for non-ram areas by checking via pat_pagerange_is_ram(). It is important that it only be called on the actual identity map because there *IS* no map to sync for highmem pages, or for memory holes.

The ioremap.c uses are not as careful as those from pat.c, and call kernel_map_sync_memtype() on PCI space which is in the middle of the kernel identity map _range_, but is not actually mapped.

Quoted 9 lines

This patch adds a check to kernel_map_sync_memtype() which probably duplicates some of the checks already in pat.c. But, it is necessary for the ioremap.c uses and shouldn't hurt other callers.

I have reproduced this bug and this patch fixes it for me and the original bug reporter:

	https://lkml.org/lkml/2013/2/5/396

Excuse me, but I didn't realize that the link is wrong.

https://lkml.org/lkml/2013/2/5/396 is a bug in CONFIG_MICROCODE_INTEL_EARLY=y && CONFIG_64BIT=n && CONFIG_DEBUG_VIRTUAL=y where patches are not available.

https://lkml.org/lkml/2013/3/5/314 is a bug in CONFIG_ACPI=y && CONFIG_DEBUG_VIRTUAL=y where your patch fixes.

Please use https://lkml.org/lkml/2013/3/5/314 rather than
https://lkml.org/lkml/2013/2/5/396 in your patch.

Regards.

#3 lore raw
Quoted 12 lines · Tetsuo Handa

Excuse me, but I didn't realize that the link is wrong.

https://lkml.org/lkml/2013/2/5/396 is a bug in CONFIG_MICROCODE_INTEL_EARLY=y && CONFIG_64BIT=n && CONFIG_DEBUG_VIRTUAL=y where patches are not available.

https://lkml.org/lkml/2013/3/5/314 is a bug in CONFIG_ACPI=y && CONFIG_DEBUG_VIRTUAL=y where your patch fixes.

Please use https://lkml.org/lkml/2013/3/5/314 rather than
https://lkml.org/lkml/2013/2/5/396 in your patch.

*Grump* I of course already committed it and other things on top.

We shouldn't really use lkml.org links anyway since if that site disappears those numbers are meaningless; rather, using Message-IDs (e.g. as embedded in lkml.kernel.org links) is better...

-hpa

maintainer
#4 lore raw
In reply to Dave Hansen

Hey Dave,

Quoted 4 lines · Dave Hansen

The original bug reporter says this fixes it for him, so I'm broadening the cc list a bit. I assume this should just get sucked in to the x86 tree.

looks like we haven't whacked all the moles - I keep seeing this when testing 32-bit builds in qemu on latest Linus + tip. I'd guess this is still that /dev/mem accessing thing called wdm.

I'm still wondering though whether we should BUG_ON on a /dev/mem access?

I've added debug output to show why we're triggering:

[  471.102902] slow_virt_to_phys((void *)x): 0x0, phys_addr: 0x37bfe000
[  471.119500] ------------[ cut here ]------------
[  471.119500] kernel BUG at arch/x86/mm/physaddr.c:85!
[  471.119500] invalid opcode: 0000 [#1] PREEMPT SMP 
[  471.119500] Modules linked in:
[  471.119500] Pid: 1571, comm: wdm Not tainted 3.9.0-rc5+ #42 Bochs Bochs
[  471.119500] EIP: 0060:[<c1032f56>] EFLAGS: 00000206 CPU: 0
[  471.119500] EIP is at __phys_addr+0x86/0xb0
[  471.119500] EAX: 37bfe000 EBX: 37bfe000 ECX: 00000001 EDX: 37bfe000
[  471.119500] ESI: f7bfe000 EDI: 00002000 EBP: f67f1f3c ESP: f67f1f28
[  471.119500]  DS: 007b ES: 007b FS: 00d8 GS: 0033 SS: 0068
[  471.119500] CR0: 8005003b CR2: bfeb12d4 CR3: 35edd000 CR4: 000006f0
[  471.119500] DR0: 00000000 DR1: 00000000 DR2: 00000000 DR3: 00000000
[  471.119500] DR6: 00000000 DR7: 00000000
[  471.119500] Process wdm (pid: 1571, ti=f67f0000 task=f5c90000 task.ti=f67f0000)
[  471.119500] Stack:
[  471.119500]  c16b1f2c 00000000 37bfe000 00000000 00002000 f67f1f64 c131d074 00002000
[  471.119500]  00000246 bfeb12ec 00000000 00000000 f67a6c40 c131d040 00002000 f67f1f8c
[  471.119500]  c1129c85 f67f1f98 f67f0000 f5ebe864 c131d040 00000020 f67a6c40 00000000
[  471.119500] Call Trace:
[  471.119500]  [<c131d074>] read_mem+0x34/0x100
[  471.119500]  [<c131d040>] ? write_mem+0x110/0x110
[  471.119500]  [<c1129c85>] vfs_read+0x85/0x130
[  471.119500]  [<c131d040>] ? write_mem+0x110/0x110
[  471.119500]  [<c1129e87>] sys_read+0x47/0xa0
[  471.119500]  [<c1546e5e>] sysenter_do_call+0x12/0x36
[  471.119500] Code: 0b a1 88 ae 0c c2 05 00 00 80 00 39 c6 72 bb a1 ac 1a 76 c1 2d 00 a0 3e 00 25 00 00 e0 ff 2d 00 20 00 00 39 c6 73 a3 0f 0b 0f 0b <0f> 0b 89 f0 e8 41 ca ff ff 89 5c 24 08 89 44 24 04 c7 04 24 2c
[  471.119500] EIP: [<c1032f56>] __phys_addr+0x86/0xb0 SS:ESP 0068:f67f1f28
[  471.508967] ---[ end trace 5fc00ac35d61284a ]---

Hmmm.

Signature · 5 lines
Regards/Gruss,
    Boris.

Sent from a fat crate under my desk. Formatting is fine.
--
#5 lore raw
Quoted 8 lines · Borislav Petkov

looks like we haven't whacked all the moles - I keep seeing this when testing 32-bit builds in qemu on latest Linus + tip. I'd guess this is still that /dev/mem accessing thing called wdm.

I'm still wondering though whether we should BUG_ON on a /dev/mem access?

We shouldn't, no. /dev/mem really needs to be fixed along a bunch of axes. Yes, it is privileged and extra creepy, but it should either work or it should fail cleanly.

-hpa

maintainer
#6 lore raw
Quoted 3 lines · H. Peter Anvin

We shouldn't, no. /dev/mem really needs to be fixed along a bunch of axes. Yes, it is privileged and extra creepy, but it should either work or it should fail cleanly.

Can't we filter out accesses through /dev/mem and not BUG_ON in __phys_addr() if they come through /dev/mem? Simply fail cleanly. No idea whether we'll break something still reading /dev/mem though.

Signature · 5 lines
Regards/Gruss,
    Boris.

Sent from a fat crate under my desk. Formatting is fine.
--
#7 lore raw
In reply to H. Peter Anvin
Quoted 11 lines · H. Peter Anvin
Quoted 6 lines · Borislav Petkov

looks like we haven't whacked all the moles - I keep seeing this when testing 32-bit builds in qemu on latest Linus + tip. I'd guess this is still that /dev/mem accessing thing called wdm.

I'm still wondering though whether we should BUG_ON on a /dev/mem access?

We shouldn't, no. /dev/mem really needs to be fixed along a bunch of axes. Yes, it is privileged and extra creepy, but it should either work or it should fail cleanly.

I've got a set sitting around collecting dust that I've got to post. It should at least restore sanity for /dev/mem on x86. I'll have it out for some initial review tomorrow.